What does "open weights" actually mean when you can't use the model for any commercial purpose?
MiniMax M2.7, a new large language model, launched with its weights publicly downloadable - meaning anyone can run the model on their own hardware. But the license blocks all commercial use without prior written approval from MiniMax. That covers paid services, commercial APIs, and deploying a fine-tuned version of the model for profit. Military applications are also explicitly banned.
This is the "open weights, closed license" pattern, and it keeps showing up.
What the License Actually Restricts
MiniMax's definition of "commercial" is deliberately broad. Building a product on M2.7 and charging for it requires permission. Running the model through an API you bill customers for requires permission. Fine-tuning M2.7 and selling access to that version requires permission. The only clearly allowed use is personal, non-commercial experimentation.
This isn't what open source means. The Open Source Initiative requires software to be freely usable for any purpose, including commercial. What MiniMax has released is the model weights - the trained parameters that determine how the AI responds - while maintaining commercial control through licensing. You can inspect and run the model, but you can't build a business on it without going back to MiniMax first.
The Practical Problem for Builders
Developers who choose open models over closed ones like Claude or ChatGPT typically do so for two reasons: cost control and legal clarity. A model released under Apache 2.0 - like many of Mistral's releases - gives you a clear license to build commercially from day one. An "open weights" model with restrictive commercial terms gives you neither advantage. You get the operational complexity of running your own infrastructure with the legal uncertainty of a proprietary model.
MiniMax isn't alone here. Meta's Llama 3.1 license blocks companies with over 700 million monthly active users from using the model at all. These restrictions matter because they determine which models developers can actually build products on long-term.
For researchers and hobbyists, M2.7 may be genuinely useful - run it locally, test its capabilities, use it for academic work. For anyone building something they plan to charge for, the license makes it a non-starter until MiniMax decides whether to open the commercial terms or keep requiring case-by-case approval.